"OpenAI already retired SWE-bench Verified, one of the most important benchmarks in AI, after 59.4% of its test cases turned out to be flawed and frontier models had memorized the answers."